From 2dc40eb21e5f439ade032c1cb982f0cf79c12c3a Mon Sep 17 00:00:00 2001
From: Matthieu KUHN <bkuhnm@l0.spartan.bench.local>
Date: Fri, 1 Apr 2022 13:33:50 +0200
Subject: [PATCH] getrf_nopiv: Add a getrf nopiv algorithm using lookhaed
buffers to enforce a ring of communications and limit the number of ongoing
communication though the dependencies.
---
compute/pzgetrf_nopiv.c | 192 +++++++++++++++++++++++++++++++++++++++-
1 file changed, 190 insertions(+), 2 deletions(-)
diff --git a/compute/pzgetrf_nopiv.c b/compute/pzgetrf_nopiv.c
index fe5bdaeec..18fb33ed8 100644
--- a/compute/pzgetrf_nopiv.c
+++ b/compute/pzgetrf_nopiv.c
@@ -26,7 +26,10 @@
*/
#include "control/common.h"
-#define A(m,n) A, m, n
+#define A(m, n) A, m, n
+#define WD(m) WL, m, m
+#define WL(m, n) WL, m, n
+#define WU(m, n) WU, m, n
/**
* Parallel tile LU factorization with no pivoting - dynamic scheduling
@@ -122,10 +125,195 @@ void chameleon_pzgetrf_nopiv_generic( CHAM_desc_t *A,
RUNTIME_options_finalize(&options, chamctxt);
}
+void chameleon_pzgetrf_nopiv_ws( CHAM_desc_t *A,
+ CHAM_desc_t *WL,
+ CHAM_desc_t *WU,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
+{
+ CHAM_context_t *chamctxt;
+ RUNTIME_option_t options;
+
+ int k, m, n, ib, p, q, lp, lq;
+ int tempkm, tempkn, tempmm, tempnn;
+ int lookahead, myp, myq;
+
+ CHAMELEON_Complex64_t zone = (CHAMELEON_Complex64_t) 1.0;
+ CHAMELEON_Complex64_t mzone = (CHAMELEON_Complex64_t)-1.0;
+
+ chamctxt = chameleon_context_self();
+ if (sequence->status != CHAMELEON_SUCCESS) {
+ return;
+ }
+ RUNTIME_options_init(&options, chamctxt, sequence, request);
+
+ ib = CHAMELEON_IB;
+ lookahead = chamctxt->lookahead;
+ myp = A->myrank / A->q;
+ myq = A->myrank % A->q;
+
+ for (k = 0; k < chameleon_min(A->mt, A->nt); k++) {
+ RUNTIME_iteration_push(chamctxt, k);
+ lp = (k % lookahead) * A->p;
+ lq = (k % lookahead) * A->q;
+
+ tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+
+ options.priority = 2*A->nt - 2*k;
+ INSERT_TASK_zgetrf_nopiv(
+ &options,
+ tempkm, tempkn, ib, A->mb,
+ A(k, k), A->mb*k);
+
+ /**
+ * Broadcast of A(k,k) along rings in both directions
+ */
+ {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ A( k, k ),
+ WL( k, (k % A->q) + lq ) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ A( k, k ),
+ WU( (k % A->p) + lp, k ) );
+
+ for ( q=1; q < A->q; q++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ WL( k, ((k+q-1) % A->q) + lq ),
+ WL( k, ((k+q) % A->q) + lq ) );
+ }
+
+ for ( p=1; p < A->p; p++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ WU( ((k+p-1) % A->p) + lp, k ),
+ WU( ((k+p) % A->p) + lp, k ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( k, k ) );
+
+ for (m = k+1; m < A->mt; m++) {
+
+ /* Skip the row if you are not involved with */
+ if ( m%A->p != myp ) {
+ continue;
+ }
+
+ options.priority = 2*A->nt - 2*k - m;
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+
+ assert( A->get_rankof( A, m, k ) == WU->get_rankof( WU, myp + lp, k) );
+ INSERT_TASK_ztrsm(
+ &options,
+ ChamRight, ChamUpper, ChamNoTrans, ChamNonUnit,
+ tempmm, tempkn, A->mb,
+ zone, WU( myp + lp, k ),
+ A( m, k ) );
+
+ /* Broadcast A(m,k) into temp buffers through a ring */
+ {
+ assert( A->get_rankof( A, m, k ) == WL->get_rankof( WL, m, (k % A->q) + lq) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempmm, tempkn,
+ A( m, k ),
+ WL( m, (k % A->q) + lq) );
+
+ for ( q=1; q < A->q; q++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempmm, tempkn,
+ WL( m, ((k+q-1) % A->q) + lq ),
+ WL( m, ((k+q) % A->q) + lq ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( m, k ) );
+ }
+
+ for (n = k+1; n < A->nt; n++) {
+
+ /* Skip the column if you are not involved with */
+ if ( n%A->q != myq ) {
+ continue;
+ }
+
+ tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
+ options.priority = 2*A->nt - 2*k - n;
+
+ assert( A->get_rankof( A, k, n ) == WL->get_rankof( WL, k, myq+lq) );
+ INSERT_TASK_ztrsm(
+ &options,
+ ChamLeft, ChamLower, ChamNoTrans, ChamUnit,
+ tempkm, tempnn, A->mb,
+ zone, WL( k, myq + lq ),
+ A( k, n ));
+
+ /* Broadcast A(k,n) into temp buffers through a ring */
+ {
+ assert( A->get_rankof( A, k, n ) == WU->get_rankof( WU, (k%A->p) + lp, n) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempnn,
+ A( k, n ),
+ WU( (k % A->p) + lp, n ) );
+
+ for ( p=1; p < A->p; p++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempnn,
+ WU( ((k+p-1) % A->p) + lp, n ),
+ WU( ((k+p) % A->p) + lp, n ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( k, n ) );
+
+ for (m = k+1; m < A->mt; m++) {
+
+ /* Skip the row if you are not involved with */
+ if ( m%A->p != myp ) {
+ continue;
+ }
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+ options.priority = 2*A->nt - 2*k - n - m;
+
+ assert( A->get_rankof( A, m, n ) == WL->get_rankof( WL, m, myq + lq) );
+ assert( A->get_rankof( A, m, n ) == WU->get_rankof( WU, myp + lp, n) );
+
+ INSERT_TASK_zgemm(
+ &options,
+ ChamNoTrans, ChamNoTrans,
+ tempmm, tempnn, A->mb, A->mb,
+ mzone, WL( m, myq + lq ),
+ WU( myp + lp, n ),
+ zone, A( m, n ));
+ }
+ }
+ RUNTIME_iteration_pop( chamctxt );
+ }
+
+ CHAMELEON_Desc_Flush( WL, sequence );
+ CHAMELEON_Desc_Flush( WU, sequence );
+
+ RUNTIME_options_finalize( &options, chamctxt );
+}
+
void chameleon_pzgetrf_nopiv( struct chameleon_pzgetrf_nopiv_s *ws,
CHAM_desc_t *A,
RUNTIME_sequence_t *sequence,
RUNTIME_request_t *request )
{
- chameleon_pzgetrf_nopiv_generic( A, sequence, request );
+ if ( ws && ws->use_workspace ) {
+ chameleon_pzgetrf_nopiv_ws( A, &(ws->WL), &(ws->WU), sequence, request );
+ }
+ else {
+ chameleon_pzgetrf_nopiv_generic( A, sequence, request );
+ }
}
--
GitLab